Date of Award

3-2023

Document Type

Thesis

Degree Name

Master of Science in Operations Research

Department

Department of Operational Sciences

First Advisor

Bruce A. Cox, PhD

Abstract

Neural networks have become increasingly popular in real time object detection algorithms. A major concern with these algorithms is their ability to quantify their own uncertainty, leading to many high profile failures. This research proposes three novel real time detection algorithms. The first of leveraging Bayesian convolutional neural layers producing a predictive distribution, the second leveraging predictions from previous frames, and the third model combining these two techniques together. These augmentations seek to mitigate the calibration problem of modern detection algorithms. These three models are compared to the state of the art YOLO architecture; with the strongest contending model achieving a 0.6% increase in precision for a 3.7% decrease in recall. This research also investigates and provides insights into what neural networks do under uncertainty. This research showed that on average for every 0.92% increase in the total number of annotations, above the mean, for a given class, the object detection model becomes 0.7% more likely to have a false positive for that class. Consequently this research presents insights that neural networks defer to the highest frequency classes from their training when they are unsure what the actual classification is.

AFIT Designator

AFIT-ENS-MS-23-M-133

Comments

A 12-month embargo was observed.

Approved for public release. Case number on file.

Share

COinS